Abstract: Abstract: Retrieval Augmented Generation (RAG) can effectively reduce the hallucination of Large Language Model (LLM) by integrating external knowledge, thereby improving the performance and reliability of natural language processing tasks. In order to systematically introduce RAG, the concepts, core technologies, and challenges of RAG was summarized.. Firstly, the basic concept of RAG was introduced, and the differences between RAG and long-context learning and fine-tuning were expounded, as well as the evaluation methods of RAG. Then, the core technology of RAG was introduced. On the one hand, focusing on the important components of RAG: retriever and generator, the key technologies were analyzed and discussed. On the other hand, the key technologies of RAG training were summarized from different angles, including independent training, joint training and agent fine-tuning. Finally, the challenges faced by RAG , such as retrieval efficiency and multimodal fusion, were discussed, and its future development direction were prospected. Despite the challenges, RAG has great potential in question answering system, code generation and text generation. Future research will focus on optimizing the performance of RAG and promoting the continuous progress of this technology.

Key words: Keywords:Large Language Model(LLM), Retrieval Augmented Generation(RAG), generative artificial intelligence, Natural Language Processing(NLP), agent

摘要： 摘 要:检索增强生成（RAG）通过结合外部知识，能有效减少大语言模型（LLM）的幻觉现象，提升自然语言处理任务的性能和可靠性。为了系统地介绍RAG，本文综述了RAG的概念、核心技术与面临的挑战。首先，介绍了RAG的基本概念，并阐述了RAG与长上下文学习、微调的区别，以及RAG的评测方法。然后，介绍了RAG的核心技术。一方面，对RAG的重要组件：检索器和生成器，进行技术分析和讨论。另一方面，从不同角度总结了RAG训练的关键技术，包括独立训练、联合训练以及agent微调。最后，探讨了RAG技术面临的挑战，如检索效率问题与多模态融合问题，并对其未来发展方向进行了展望。尽管存在挑战，但RAG在问答系统、代码生成及文本生成等领域具有巨大潜力。未来研究将致力于优化RAG系统性能，推动该技术的持续进步。

关键词: 大语言模型, 增强检索生成, 生成式人工智能, 自然语言处理, 智能体